Source-linked AI summary
DetNet: A Backbone network for Object Detection
Zeming Li, Chao Peng, Gang Yu, Xiangyu Zhang, Yangdong Deng, Jian Sun
TL;DR
Object detectors commonly rely on ImageNet classification backbones even though detection requires detector-specific stages and precise spatial localization. DetNet addresses this gap with a backbone that adds stages while preserving high spatial resolution and receptive field, and it reports strong COCO detection and instance-segmentation results.
Problem
Object detectors commonly reuse ImageNet classification backbones although detection requires additional scale-handling stages and spatial localization.
Method
DetNet is a detection-specific backbone that matches detector stages, preserves deeper-layer spatial resolution, and uses low-complexity dilated bottlenecks.
Results
DetNet reports impressive results for object detection and instance segmentation on the COCO benchmark.
Takeaways & Limitations
The paper supports designing backbone networks for detection rather than directly transferring classification-oriented backbone structures.
Takeaways & Limitations
DetNet-59 has more parameters than ResNet-50 because it includes an additional FPN P6 stage.
Abstract
from arXiv · showhide
Recent CNN based object detectors, no matter one-stage methods like YOLO, SSD, and RetinaNe or two-stage detectors like Faster R-CNN, R-FCN and FPN are usually trying to directly finetune from ImageNet pre-trained models designed for image classification. There has been little work discussing on the backbone feature extractor specifically designed for the object detection. More importantly, there are several differences between the tasks of image classification and object detection. 1. Recent object detectors like FPN and RetinaNet usually involve extra stages against the task of image classification to handle the objects with various scales. 2. Object detection not only needs to recognize the category of the object instances but also spatially locate the position. Large downsampling factor brings large valid receptive field, which is good for image classification but compromises the object location ability. Due to the gap between the image classification and object detection, we propose DetNet in this paper, which is a novel backbone network specifically designed for object detection. Moreover, DetNet includes the extra stages against traditional backbone network for image classification, while maintains high spatial resolution in deeper layers. Without any bells and whistles, state-of-the-art results have been obtained for both object detection and instance segmentation on the MSCOCO benchmark based on our DetNet~(4.8G FLOPs) backbone. The code will be released for the reproduction.
1 Introduction
DetNet is proposed because standard object detectors commonly fine-tune ImageNet classification backbones despite differences in stage requirements and spatial localization. The backbone adds detector-aligned stages while preserving high-resolution features and a large receptive field.
- Recent one-stage and two-stage CNN detectors commonly depend on backbones pretrained for ImageNet classification.
- Classification backbones create a task gap because object detection requires both category recognition and spatial bounding-box localization.
- DetNet introduces stages needed for varying object scales while maintaining feature-map spatial resolution in deeper layers.
- A low-complexity dilated bottleneck helps DetNet retain high-resolution feature maps and a large receptive field.
- DetNet is reported to achieve new state-of-the-art results on MSCOCO object detection and instance segmentation using DetNet59.
2 Related Works
Prior work generally builds detection systems on ImageNet classification backbones, while the suitable backbone for object detection remains underexplored. DetNet is introduced as a backbone designed specifically for detection rather than directly following that classification-oriented pipeline.
- 2.1 Backbone Network: Object-detection backbones are usually borrowed from networks designed for ImageNet classification.
- 2.1 Backbone Network: Classification networks commonly use repeated downsampling and 32-stride feature maps to reduce computation and increase receptive field.
- 2.2 Object Detection Business Part: One-stage detectors use backbones for direct instance prediction, while SSD and FPN-style designs exploit multi-layer or pyramidal features for varying object scales.
- 2.2 Object Detection Business Part: The suitable backbone for object detection is described as an unexplored field because recent detectors generally follow ImageNet-pretrained pipelines.
- 2.2 Object Detection Business Part: DetNet addresses this gap by designing a backbone specifically for object detection.
3 DetNet: A Backbone network for Object Detection
DetNet aligns backbone stages with detector requirements while preserving high spatial resolution and receptive field. Its design targets localization weaknesses of classification backbones and manages the computational cost of deeper high-resolution processing.
- 3.1 Motivation: Detectors add stages such as FPN P6, but these stages are not pretrained in ImageNet classification networks.
- 3.1 Motivation: Large strides can blur large-object boundaries and weaken or remove small-object information from deeper feature maps.
- 3.1 Motivation: DetNet directly designs the number of stages for detection and maintains high-resolution feature maps despite using six or seven stages.
- 3.1 Motivation: DetNet matches the detector’s stage count, allowing extra stages such as P6 to be pretrained, while its final-stage resolution supports large-object localization and small-object discovery.
- 3.2 DetNet Design: DetNet keeps stages 1–4 identical to ResNet-50 and begins its architectural changes at stage 5.
- 3.2 DetNet Design: After stage 4, DetNet fixes resolution at 16x downsampling and introduces stages using dilated bottlenecks with 1x1 convolution projections.
- 3.2 DetNet Design: Dilated bottlenecks enlarge receptive fields efficiently, while stages 5 and 6 retain stage-4 channel width to control cost.
- 3.2 DetNet Design: DetNet can be integrated with detectors with or without feature pyramids, and FPN is used as the baseline while other structures remain fixed.
4 Experiments
Experiments evaluate DetNet as an object-detection backbone across COCO detection and instance segmentation, including controlled comparisons with ResNet and ablations of its stage design. DetNet improves detection accuracy, especially large-object localization and small-object recall, while maintaining lower complexity than a stronger ResNet baseline.
- Main Results: DetNet-59 replaces ResNet-50 in FPN while preserving the detector’s other structures for backbone evaluation.The baseline is FPN with ResNet-50, and DetNet-59 adds an additional stage.
- Main Results: Over 2 points in mAP: DetNet-59 outperforms ResNet-50 in FPN despite using 4.8G FLOPs.DetNet-59 has 23.5% ImageNet top-1 error; ResNet-101 uses 7.6G FLOPs and reaches 39.8 mAP, below DetNet-59.
- Main Results: 1.8 points: DetNet-59 still outperforms ResNet-50 when both FPN backbones are trained from scratch.The experiment uses synchronized batch normalization during scratch training.
- Results Analysis: 5.5 points in AP85@large: DetNet-59 improves large-object localization from 34.5 to 40.0.The authors attribute this to avoiding overly large strides in deeper feature maps, which can blur large objects for accurate regression.
5 Conclusion
The paper presents DetNet as a backbone designed specifically for object detection, achieving state-of-the-art results on MSCOCO object detection and instance segmentation.
- DetNet is a novel backbone network designed specifically for the object detection task.
- DetNet-59 outperforms previous state-of-the-art methods on MSCOCO test-dev object detection with much lower FLOPs.
- DetNet-59 achieves a new state-of-the-art result on the MSCOCO test-dev instance segmentation task.
- The reported results demonstrate strong performance for both object detection and instance segmentation on the COCO benchmark.